Introduction: Currently, most of invasive brain-machine interfaces (BMI) rely on signals recorded using electrodes implanted intra-cortically or subdural electrocorticography (ECoG) arrays. Barring a few studies [1,2], epidural electrodes–often used for chronic stimulation to alleviate neuropathic pain—are seldom used for this purpose. Here we report their use to successfully control a brain-machine interface over several days. Methods: Experiments were performed with a 50 years old male patient who suffered a left brachial plexus avulsion 30 years prior to the surgery and recordings leading to a complete left arm plegia. Two epidural leads (4 channels each) were implanted above central sulcus on the primary motor and sensory cortex contralateral to the plegic hand to apply epidural stimulation to treat deafferentation pain. Contact leads were temporarily externalized for 9 days allowing recording and decoding of cortical activity during attempted movements. Five experimental sessions were performed (1, 3, 4, 8 and 9 days after the operation). The subject was asked to attempt to move his plegic hand as if controlling the cursor with a mouse towards one target location in a screen. After a cue shows the target the subject should wait until it becomes green before starting the movement, then stopping once the target is reached (Fig 1 Left). Each session lasted less than 2 hours and yielded about 80 trials. ECoG signals were recorded at a sampling rate of 512 Hz (8 channels corresponding to the 2 implanted leads; reference and ground electrodes located at the two mastoids). Data was processed in real time by extracting the spectral power in the range of 2-40 Hz. Data from the first day was used to train an initial classifier to discriminate between resting and movement periods. From the second day onwards, the classifier output was used to control the movement of the cursor. Whenever the classifier identified the neural activity as corresponding to the movement, the cursor was displaced towards the target location. Before each session, the classifier was updated using the data from previous sessions to assess the stability of the decoder.